> The Recall Ceiling: Why LLM Reranking Cannot Fix Sparse Retrieval

Recommender Systems 高级 2026-04-01 06:41 2026-04-01
#recall ceiling #LLM reranking #DPI bound #retrieval #information theory #recommendation

A formal DPI-derived bound proves reranking cannot recover items absent from retrieval. Across 8 datasets: 73-91% zero-recall users make even perfect reranking yield near-zero NDCG. Adaptive-K raises the ceiling by 62-105%.

The Recall Ceiling

The Theorem (Data Processing Inequality)

Any reranker operating on K candidates cannot produce items outside that set. At typical R@100 = 5% with 10 relevant items per user, max NDCG@10 ≈ 0.09 regardless of reranker quality.

8-Dataset Evidence

Dataset R@100 Zero-Recall Users Best Ceiling Utilization
Beauty 8.4% 88% 10%
Movies 3.6% 73% 12%
Electronics 2.8% 80% 7%
Toys 13.8% 91% 53% (LambdaMART)
MovieLens 18.9% 46% 10%

LambdaMART (2011) achieves best ceiling utilization — not any modern deep learning reranker.

LLM Reranking Result

Qwen2.5-7B ceiling utilization: only 25-32%. 68% of potential gain is unrealized even within the constrained recall set.

Counterintuitive Warm User Finding

Warm users (more history) have lower retrieval recall. They've consumed popular items, leaving long-tail test items that CF cannot retrieve.

Adaptive-K: The Fix

Dataset Fixed R@K Adaptive R@K Gain
Beauty 8.2% 13.9% +70%
Movies 3.0% 6.1% +105%
MovieLens 18.9% 36.0% +91%

Per-user candidate set sizing based on estimated sparsity raises the ceiling itself.

Practical Rule

Before investing in LLM reranking: compute catalog density. Dense domains (MovieLens: 1.96%) make reranking worthwhile. Sparse domains (Movies: 0.27%) make even perfect reranking yield negligible gains.